高斯过程回归在线软测量建模改进研究

发布时间：2018-04-26 22:11

本文选题：高斯过程回归 + 模型更新　；参考：《江南大学》2017年硕士论文

【摘要】：实际工业过程多具有非线性、时变性及不确定性等特点,而传统离线软测量模型无法对此类工业过程的状态参数进行实时跟踪。针对上述问题,通常对传统离线软测量方法进行自适应改进,并根据实时数据对在线软测量模型参数及数据库进行一定的预处理和更新,以确保所建软测量模型具备跟踪过程动态特征及抗干扰的能力,最终进一步提高模型精度以及性能。为了实现对过程主导变量进行有效预测及控制,本文首先采用高斯过程回归(Gaussian Process Regression,GPR)算法对实际工业过程进行学习并得到相应软测量模型,随后提出动态模型更新及奇异点检测补偿算法对其进行动态校正和预处理,最终通过实验仿真论证了本文所提算法的有效性。论文的主要研究内容如下所示:(1)高斯过程回归算法实际应用研究。首先对高斯过程回归算法原理进行简要解析。随后利用此回归算法对青霉素发酵过程进行学习并建立相应软测量模型。通过与传统最小二乘支持向量机(Least Squares Support Vector Machine,LSSVM)的仿真对比,表明所建高斯过程回归模型具有更好的预测性能。(2)进一步考虑工业过程的时变特征,提出一种基于动态模型更新的GPR在线软测量方法。该方法首先对训练样本利用GPR方法进行离线建模,得到预测输出及预测误差;然后对所得预测误差进行分析,当误差均值大于某一预设阈值时对GPR模型进行整体更新:同时更新其协方差矩阵和协方差函数的参数;否则,只对GPR模型进行局部更新:只更新其协方差矩阵。接着利用误差高斯混合模型(Error Gaussian Mixture Model,EGMM)对更新后的GPR模型预测输出进行误差补偿从而得到最优的预测结果。最终由实际工业污水处理过程的实例仿真验证了所提方法的有效性。(3)针对软测量方法在实际应用中查询样本可能出现奇异点这一问题,提出一种带奇异点检测补偿的GPR在线软测量方法。该方法首先对训练样本利用GPR方法进行建模,得到软测量模型;然后对新来查询样本采用改进拉依达准则进行奇异点检测,当新来查询样本被确定为奇异点时,利用辅助模型对奇异点进行修补,然后再利用软测量模型对修补后查询样本点进行预测;否则,直接使用软测量模型对新来查询样本点进行预测。最终通过实际硫回收过程数据的实验仿真验证了所提方法的有效性。
[Abstract]:The actual industrial processes are usually nonlinear, time-varying and uncertain, but the traditional off-line soft sensor model can not track the state parameters of such industrial processes in real time. In order to solve the above problems, the traditional off-line soft sensing methods are usually improved adaptively, and the parameters of the online soft-sensing model and the database are preprocessed and updated according to the real-time data. In order to ensure the dynamic characteristics of tracking process and the ability of anti-jamming, the model can improve the precision and performance of the model. In order to effectively predict and control the process leading variables, this paper first uses Gao Si process regression Gaussian Process algorithm to study the actual industrial process and obtains the corresponding soft sensor model. Then a dynamic model updating and singular point detection compensation algorithm is proposed to dynamically correct and preprocess it. Finally, the effectiveness of the proposed algorithm is demonstrated by experimental simulation. The main contents of this paper are as follows: 1) Gao Si process regression algorithm. Firstly, the principle of Gao Si process regression algorithm is analyzed briefly. Then the regression algorithm was used to study the penicillin fermentation process and the corresponding soft sensor model was established. Compared with the traditional least square support vector machine (LSSVM), it is shown that the proposed Gao Si process regression model has better predictive performance and further considers the time-varying characteristics of industrial processes. An online soft sensor method for GPR based on dynamic model updating is proposed. In this method, the training sample is modeled off-line by GPR method, and the prediction output and prediction error are obtained, and then the prediction error is analyzed. When the mean error is greater than a preset threshold, the GPR model is updated as a whole: the covariance matrix and the parameters of the covariance function are updated at the same time; otherwise, the GPR model is only locally updated: only its covariance matrix is updated. Then the error Gaussian Mixture model EGMMM is used to compensate the error of the updated GPR model to get the best prediction result. Finally, the effectiveness of the proposed method is verified by a practical example of industrial wastewater treatment. 3) aiming at the problem that the sample may appear singularity in the application of soft sensor, This paper presents an online soft sensing method for GPR with singularity detection compensation. In this method, the training sample is modeled by GPR method, and the soft sensor model is obtained, and then the new query sample is detected by the improved Laida criterion, when the new query sample is determined as the singularity point. The singular points are repaired by the auxiliary model, and then the sample points are predicted by the soft sensor model. Otherwise, the new query sample points are predicted directly by the soft sensor model. Finally, the effectiveness of the proposed method is verified by the experimental simulation of the actual sulfur recovery process data.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TQ018

【参考文献】